[Uneadlight]

What

TokenLogprob now keeps the bytes field the server already sends.

Why

With logprobs=True, /api/generate and /api/chat return a bytes array (the
token's raw UTF-8 bytes) on every entry of logprobs and top_logprobs:

{"token": "Hello", "logprob": -0.20118224620819092, "bytes": [72, 101, 108, 108, 111]}

TokenLogprob only declared token and logprob, and since
SubscriptableBaseModel does not set extra='allow', pydantic's default
extra='ignore' dropped the field without any error:

>>> r = client.generate(model='qwen2.5:3b', prompt='Say hi in three words.', logprobs=True)
>>> r.logprobs[0].model_dump()
{'logprob': -0.2524, 'token': 'Hello', 'top_logprobs': None}
>>> r.logprobs[0].bytes
AttributeError: 'Logprob' object has no attribute 'bytes'

After the change:

>>> r.logprobs[0].bytes
[72, 101, 108, 108, 111]

How

Add one optional field to TokenLogprob (Logprob inherits it):

bytes: Optional[Sequence[int]] = None
'Raw UTF-8 bytes of the token. Useful when a token is not a standalone valid string.'

It is kept as a list of ints rather than a bytes object so model_dump()
stays JSON-serializable.

Older servers that do not send the field still parse: the attribute is None and
model_dump(exclude_none=True) omits it, which is the backward-compatible
behavior requested in #750.

Tests

4 added to tests/test_type_serialization.py:

• test_generate_response_keeps_logprob_bytes
• test_chat_response_keeps_logprob_bytes
• test_logprob_bytes_is_optional
• test_top_logprobs_keep_bytes

pytest tests: 113 passed, 8 failed. The 8 failures are pre-existing and
unrelated (test_client_create_blob* hits PermissionError on this machine and
the image serialization tests fail with Invalid image data due to a broken
%TEMP%); the same 8 fail on main without this change.

ruff check and ruff format --check: clean.

Fixes #750